Comparison of chemometrics strategies for the spectroscopic monitoring of active pharmaceutical ingredients in chemical reactions

The Process Analytical Technology (PAT) initiative promoted by the Food and Drug Administration (FDA) encourages pharmaceutical companies to increase the use of new analytical technologies to perform constant monitoring of the critical quality attributes (CQA), allowing a better understanding and a better control of the process. This paper presents a practical framework based on different dimension-reduction methods as well as calibration methods aimed at following over time chemical experiments organized in batches. To illustrate it, this paper uses pharmaceutical data collected in a research and development context towards industrial production. This methodological framework aims to reach two objectives. The first objective is to visualize and interpret in real time, or off-line, the kinetics of chemical reactions using the following dimension-reduction methods: principal component analysis (PCA), non-negative matrix factorization (NMF) and multivariate curve resolution (MCR). The results show that, due to their additional constraints, NMF and MCR allow a better interpretability of chemical reactions than PCA with a comparable quality of fit. Moreover, eventough NMF and MCR come from different fields, their algorithms share many similarities and produce close results. The second objective is to predict chemical component concentrations over time. For this second objective, the partial least squares regression (PLSR) is used in a one-step approach and compared with a two-step approach combining multivariate regression with PCA, NMF or MCR. The results show that spectra or scores obtained from unsupervised approaches PCA, NMF or MCR can be used to predict concentrations of the main chemical compounds continuously over all the time of the reaction with a good precision and with a gain of interpretability. For both objectives, possible model validation indices are also discussed including a leave-one-batch-out approach.